Multilead ECG Delineation Using Spatially Projected Leads From Wavelet Transform Loops

1996 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 8, AUGUST 2009

Multilead ECG Delineation Using Spatially ProjectedLeads From Wavelet Transform Loops

Rute Almeida∗, Juan Pablo Martınez, Ana Paula Rocha, Member, IEEE,and Pablo Laguna, Senior Member, IEEE

Abstract—In this paper, a novel multilead (ML) based automaticstrategy for delineation of ECG boundaries is proposed and eval-uated with respect to the QRS and T-wave boundaries. The MLstrategy is designed from a single-lead (SL) wavelet-transform-based delineation system. It departs from three orthogonal leadsand takes advantage of the spatial information provided using aderived lead better fitted for delineation. SL delineation is thenapplied over this optimal derived lead. The ML strategy producesa reduced error dispersion compared to SL results, thus provid-ing more robust, accurate, and stable boundary locations than anyelectrocardiographic lead by itself and outperforming strategiesbased on lead selection rules after SL delineation.

Index Terms—Derived lead, loop, multilead (ML) ECG, vecto-cardiogram (VCG), wave delineation, wavelet transform (WT).

I. INTRODUCTION

THE DIFFERENT phases of the heart’s electrical activityare mapped to the waves in the ECG, typically known as P,

Q, R, S, and T waves. The delineation of the ECG characteristicwaves in each cardiac beat consists of detecting their peaks andboundaries (onset and end), and it also provides fundamentalfeatures to derive clinically useful information, namely aboutthe duration of the electrical phenomena and their beat-to-beatevolution.

The detection of the QRS complex is the first and the moststraightforward stage of any delineation system. The beat lo-cation is defined by the mark of the QRS complex main wave

Manuscript received October 27, 2008; revised February 20, 2009 and April 2,2009. First published May 12, 2009; current version published July 15, 2009.This work was supported in part by project TEC2007-68076-c02-02 from theMinistry of Science and Innovation and European Regional Development Fund(FEDER), Grupo Consolidado Communications Technology Group (GTC) fromDiputacion General de Aragon (DGA) T:30, and Centro de Matematica da Uni-versidade do Porto (CMUP), financed by Foundation for Science and Tech-nology, from the portuguese Fundacao para a Ciencia e a Tecnologia (FCT),Portugal, through the programmes POCTI and POCI 2010, with national and Eu-ropean Community Structural Funds. Asterisk indicates corresponding author.

∗R. Almeida is with the Centro de Investigacion Biomedica en Red enBioingenierıa, Biomateriales y Nanomedicina (CIBER-BBN), 50018 Zaragoza,Spain, and also with the Communications Technology Group (GTC), AragonInstitute of Engineering Research (I3A), University of Zaragoza, 50018Zaragoza, Spain. He is also with the Centro de Matematica da Universidadedo Porto (CMUP), Rua Campo Alegra 687, 4169-007 Porto, Portugal (e-mail:[email protected]).

J. P. Martınez and P. Laguna are with the Communications Technology Group(GTC), Aragon Institute of Engineering Research (I3A), University of Zaragoza,50018 Zaragoza, Spain, and also with the Centro de Investigacion Biomedicaen Red en Bioingenierıa, Biomateriales y Nanomedicina (CIBER-BBN), 50018Zaragoza, Spain (e-mail: [email protected]; [email protected]).

A. P. Rocha is with the Departamento de Matematica Aplicada, Faculdadede Ciencias da Universidade do Porto, 4169-007 Porto, Portugal, and also withthe Centro de Matematica da Universidade do Porto (CMUP), 4169-007 Porto,Portugal (e-mail: [email protected]).

Digital Object Identifier 10.1109/TBME.2009.2021658

(usually the R wave), and the search for peaks and boundaries ofthe ECG waves is usually performed within temporal windowsreferred to the QRS position. Especially problematic due tothe low SNR is the delineation of low-amplitude smooth waveboundaries, as it is usually the case of the T-wave end. Fur-thermore, there are no standard clear rules to locate the waves’boundaries, and this makes systematizing the delineation moredifficult. Automatic methodologies allow one to avoid intra-/interobserver variability, and therefore, developing accurate androbust methods for ECG automatic delineation is a topic of maininterest. A wide diversity of algorithms for QRS detection anddifferent delineation approaches have been proposed, regardingsome or all of the ECG waves and limits. In particular, an au-tomatic single-lead (SL) delineation system, which generalizesthe wavelet transform (WT) based methodology of [1], was pre-sented in [2]. The method in [2] is able to determine and locatethe peaks and boundaries of the QRS complex (namely identify-ing its individual waves), and of P and T waves. We refer to theability of the algorithm to deal with various P and T wave mor-phologies, namely positive, inverted, biphasic, accounting fordifferent morphologies. The WT provides a description of thesignal in the time-scale domain, thereby allowing the represen-tation of its temporal features at different resolutions (scales) ac-cording to their frequency content. Thus, regarding the purposeof locating different waves with typical frequency characteris-tics, the WT is a suitable tool for ECG automatic delineation.

According to the dipolar hypothesis, the electrical activityof the heart can be approximated by a time-variant electricaldipole, called the electrical heart vector (EHV). Thus, the volt-age measured at a given lead would be the projection of theEHV into the unitary vector defined by the lead axis [3]. Choos-ing a particular lead for ECG delineation determines a point ofview over the cardiac phenomena, and different latencies on thewaves’ onsets and ends are found in different leads. Neverthe-less, the onset and end of the cardiac electric phenomena areindeed unique, and therefore, a global feature for all the leads.Thus, adequately combining the information provided by multi-ple leads is essential for the correct location of lead-independentwaves’ boundaries. The SL system [2] includes postprocessingdecision rules to deal with multilead (ML) files by choosingglobal marks based on the SL-based sets of locations. However,this system is not truly ML, and it requires to apply the SLmethodology to a large number of leads.

In this paper, an actual ML methodology regarding bound-aries location is proposed and validated. The ML approach de-parts from the SL system and attends to the spatial character-istics of the different available leads, aiming to achieve a morerobust delineation. This ML system was partially validated and

0018-9294/$25.00 © 2009 IEEE

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on July 22, 2009 at 04:30 from IEEE Xplore. Restrictions apply.

ALMEIDA et al.: MULTILEAD ECG DELINEATION USING SPATIALLY PROJECTED LEADS FROM WAVELET TRANSFORM LOOPS 1997

Fig. 1. Example of SL delineation of T-wave boundaries: nT denotes T-wavepeak (zero crossing of the WT), nf and nl the first and last significant WTslopes associated with the T wave, no and ne the T-wave onset and end, and ξo

and ξe the thresholds for T-wave boundaries. Note that T-wave end is markedusing criterion 1) while T-wave onset satisfies criterion 2).

compared with other approaches as part of the “PhysioNet/computers in cardiology challenge 2006: QT interval measure-ment” competition [4].

II. MATERIALS AND METHODS

A. SL Delineation

The SL-based delineation system is described in detail else-where [2], and only general features are referred to here. Thediscrete dyadic WT is implemented using the algorithm atrousto maintain temporal resolution at different scales. The wavelet-equivalent filter for each scale is resampled to keep the fre-quency band response independent of the sampling frequencyof the ECG signal. The detection of the fiducial points is carriedout across the adequate WT scales, attending to the dominantfrequency components of each ECG wave: QRS waves corre-spond to a simultaneous effect in scales 21–24 , while the T andP waves affect mainly scales 24 or 25 . The prototype waveletused (a derivative of a smoothing function) allows to obtain aWT at scale 2m , wx,m [n], which is proportional to the deriva-tive of the filtered version of the signal x[n] with a smoothingimpulse response at scale 2m . Thus, ECG wave peaks corre-spond to zero crossings in the WT, and ECG maximum slopescorrespond to WT’s maxima and minima.

Depending on the number and polarity of the slopes found,a wave morphology is assigned and boundaries are located us-ing threshold-based criteria. The onset (end) of a wave, no(ne),occurs before (after) the first (last) significant slope associatedwith the wave (the first (last) maximum of |wx,m [n]|), at sam-ple nf (nl) (Fig. 1). Each boundary is located by selecting thesample nearest to nf (nl) where one of the following two crite-ria is satisfied (Fig. 1): 1) |wx,m [n]| is below a threshold ξo(ξe)relative to wx,m [nf ] (wx,m [nl ]) and 2) a local minimum of|wx,m [n]| exists before nf (after nl).

B. SL Selection Rule for ML Signals

To deal with ML signals and obtain global marks for peaklocation, a median postprocessing selection rule over SL-basedlocations is used. For boundaries location, the possibly differentlatencies between leads resulting from their spatial orientationneed to be taken into account. Thus, the postprocessing rules forboundaries consist of ordering the SL annotations and selectingas the onset (end) of a wave the first (last) annotation whose knearest neighbors lay within a δ ms interval. To combine 12 SLannotations, k = 3, δ = 10 ms were used for QRS end, and δ =12 ms for QRS onset and T end [5].

C. ML Delineation

The ML delineation system proposed here considers threesimultaneous orthogonal leads: x[n], y[n], and z[n]. The vec-tocardiogram (VCG) is an EHV’s canonical representation de-fined by three orthogonal leads [3], which are usually acquiredas the corrected Frank leads, and is given by

s[n] = [x[n], y[n], z[n]]T . (1)

An example of a beat according to the Frank leads and therespective VCG spatial loop for a T wave is plotted in Fig. 2(a)and (c).

A spatial WT loop in a time window W is given by

wm [n] = [wx,m [n], wy,m [n], wz,m [n]]T (2)

for a given scale 2m∣∣m∈{1,2,3,...} , as illustrated in Fig. 2(b) and

(d). As a consequence of the WT prototype used, the WT loopwm [n] |n∈W is proportional to the VCG derivative and describesthe velocity of evolution of the EHV in a time interval W .Assuming that the noise is spatially homogeneous, the directionwith maximum projection of the WT in the region close to thewave boundary would define the ECG lead maximizing the localSNR, and thus, the most appropriate for boundary delineation.The main direction u = [uX , uY , uZ ]T of EHV variations at ascale 2m on any time interval W is given by the director vectorof the best straight linear fit to all points in the WT loop wm [n][Fig. 2(d)]. By adequately choosing the time interval W , it ispossible to find the u corresponding to the lead most suited fordelineation purposes. It should be noted that the time intervalsI (used for projecting) and W (used for linear fitting) can bedifferent, depending on each wave’s specificity.

The strategy proposed for ML boundary delineation usingWT loops is based on a multistep iterative search for a betterspatial lead for delineation improvement (with steeper slopes),particularized for each boundary. At each step (i), the vectoru(i) is determined separately for each beat and boundary, byadapting and updating the interval W , defined specifically foreach wave and boundary, in a way to increase the SNR andensure steep slopes in w

(i)d,m [n] obtained by (4). The goal is to

construct a derived wavelet signal well suited for boundaries lo-cation, using the same detection criteria as in the SL delineator.1

1SL-threshold-based criteria were applied as reported in [2] except for T-wave

end with ξe (m) = 0.25wx,m [n(i)l ].



Fig. 2. Examples of ECG and the components of wm [n] corresponding towhole beat in Frank leads (n ∈ W for the gray area) and VCG loop and WTloop corresponding to the T wave (n ∈ W ). (a) ECG leads. (b) WT at scale 24 .(c) VCG loop of the T wave. (d) w4 [n] loop of the T wave, best line fit by TLSminimization to the entire loop and some fitting residues in TLS sense.

Considering the VCG loop s[n] in any time interval I , aderived lead d[n] defined by the axis u can be constructed byprojecting the points of the VCG loop

d[n] =sT [n]u‖u‖ , n ∈ I. (3)

The variable d[n] combines the information provided by thethree leads in s[n]. Instead, the WT loop (wm [n]) can be pro-jected, and a derived wavelet signal wd,m [n], corresponding tothe ECG lead defined by the axis u, can be constructed as

wd,m [n] =wT

m [n].u‖u‖ , n ∈ I. (4)

1) General Algorithm for ML Boundary Location: For eachbeat and boundary, the following strategy is applied (particulardetails for each boundary are described in later sections).

INITIALIZATIONa0) An initial search window adequate to find the EHV’s

main direction in the boundary is defined as W (1) .b0) The initial main direction of EHV variations u(1) is es-

timated using the adequate scale 2m , as the best line fitin total least squares (TLS) sense [6] to wm [n] |n∈W ( 1 ) .

c0) The loop wm [n]∣∣n∈[nQRS , k −1 , nQRS , k + 1 ] , with nQRS,k

being the median of SL-derived locations for the QRScomplex in the kth beat, is projected over u(1) to con-struct the new derived WT signal w

(1)d,m [n].

d0) SL delineation over w(1)d,m [n] allows to locate n

(1)o or

n(1)e , the boundary position at step 1.

ITERATION—STEP (i)a) The search window W (i) is updated while attending to the

boundary location provided by the previous step (i − 1).b) The main direction of EHV variations u(i) is estimated as

the TLS best line fit to wm [n], n ∈ W (i) .c) The new derived WT signal w

(i)d,m [n] is constructed by

projecting the loop wm [n]∣∣n∈[nQRS , k −1 , nQRS , k + 1 ] .

d) IF w(i)d,m [n] is less fitted for boundary location than

w(i−1)d,m [n], OR no significant maximum of w

(i)d,m [n] was

found THEN n(i−1)o or n

(i−1)e (found in the previous step)

is adopted as ML mark; STOP; ELSE SL delineation ofthe boundary is performed over w

(i)d,m [n] to find n

(i)o or

n(i)e updated marks.

e) IF no relevant change is found in the boundary locationTHEN n

(i)o or n

(i)e is adopted as ML mark; STOP; ELSE

REPEAT from a).In the traditional least-squares criteria, it is assumed that er-

rors occur in only one observed variable, while the other vari-ables are exactly known. However, all the three signals used todefine the WT loop of (2) are representations of observed ECGleads that can contain noise contamination. The TLS fitting doesnot assume that some variables are error-free and minimizes thedistance between each observation and the fitted line (orthogo-nal deviations are considered [6]), as illustrated in Fig. 2(d). Itmust be also remarked that the use of WT loop to base the leaddirection instead of taking directly the VCG loop is relevant, asit allows to avoid the high- and low-frequency noise contam-ination, and thus, produces a more accurate lead selection fordelineation [Fig. 2(c) and (d)]. In addition, if the Frank leadswere not recorded:

1) Frank leads can also be synthesized from the standard12-lead system [7];

2) the ML delineation methodology can be applied over anyset of three ECG orthogonal leads;

3) considering WT loops in a 2-D plane instead of in a 3-Dspace is also possible, thus allowing to apply this method-ology to any two ECG orthogonal leads.

In cases of extreme noise or with very poor signal amplitude,the lead search can fail in finding the optimal projection, giving



Fig. 3. Example of ML delineation (files from CSEDB database [9]): WTloops and the direction of the best line fit at the initial and final step (i = 2).(a) ML QRS onset location. (b) ML T end location.

inconsistent boundaries (e.g., an onset after the wave peak).These boundaries were discarded.

2) Specific Parameters for QRS Complex Boundaries: MLlocation of the QRS boundaries is performed using the WT loopat scale 22 , as illustrated for QRS onset in Figs. 3(a) and 4(a). Letus define n

(0)QRS,o(n

(0)QRS,e) as the earliest (latest) QRS onset (end)

location given by the SL methods (over each orthogonal lead)and n

(0)QRS,f (n(0)

QRS,l) is the earliest (latest) significant maximummodulus location. The initial search window for QRS onset andend is taken as

W (1) = Q(1) =[n

(0)QRS,o − 4sCSE(QRSo n ) , n

(0)QRS,f

](5)

W (1) = S(1) =[n

(0)QRS,l , n

(0)QRS,e + 4sCSE(QRSe n d )

](6)

respectively, where sCSE(QRSo n ) = 3.25 ms andsCSE(QRSe n d ) = 5.8 ms are the standard deviation toler-ance values provided in [8]. At each iteration (i), the search

Fig. 4. Comparison between SL and ML delineation: ECG in orthogonal leadsx[n], y[n], z[n], the corresponding WT signals wx,m [n], wy ,m [n], wz ,m [n],the derived ECG signals d(1) [n], d(i) [n], and WT signals w

(1)d ,m

[n], w(i)d ,m

[n],

following the directions of vectors u(1) and u(i) found for the kth beat (Fig. 3),respectively, for QRS onset and T-wave end. Vertical dashed lines stands forthe mark found in the respective lead, solid line stands for median refereemarks, and the stars stands for the first (last) significant maximum modulusin the constructed lead. (a) QRS onset delineation (m = 2, n is an arbitraryneighborhood of the QRS complex for the kth beat), i = 2. (b) T end delineation(m = 4, n ∈ T w[1 ] , i = 2).

window is updated

W (i) = Q(i) =[n

(i−1)QRS,o − 4sCSE(QRSo n ) , n

(i−1)QRS,f

](7)

W (i) = S(i) =[n

(i−1)QRS,l , n

(i−1)QRS,e + 4sCSE(QRSe n d )

](8)

where n(i−1)QRS,o (n(i−1)

QRS,e) is the QRS onset (end) position found

at step (i − 1) and n(i−1)QRS,f (n(i−1)

QRS,l) is the location of the first

(last) significant maximum modulus of w(i−1)d,m [n].

IF n(i)QRS,f (n(i)

QRS,l) has the same polarity as n(i−1)QRS,f (n(i−1)

QRS,l),equal or lower amplitude, and QRS complex morphology



includes a Q (S) wave, it is considered that the lead constructedat step (i) is not better for QRS onset (end) location than thatconstructed at step (i − 1) and the iteration STOP criterion ofd) is applied. The iteration STOP criterion of e) is applied if thethe same location is achieved for three iterations.

3) Specific Parameters for T-Wave Boundaries: ML delin-eation of T-wave boundaries is illustrated for the T end inFigs. 3(b) and 4(b). Scale 2m = 25 is considered if in the SLdelineation, the scale 25 was used for T-wave detection for atleast two out of the three leads, and m = 4 otherwise [2]. TheT-wave morphology is typically more simple; thus, a single ini-tial search window W (1) = T(1)

o = T(1)e , taken as the union of

the SL search windows on the orthogonal leads, is consideredfor each beat, both regarding the wave’s onset and end. Thesearch window for T onset at iteration (i) is updated as

W (i) = T(i)o =

[n

(i−1)T ,o − 4sCSE(Ton); n

(i−1)T ,f

]

and the search window for T end is actualized as

W (i) = T(i)e =

[n

(i−1)T ,l ; n

(i−1)T ,e + 4sCSE(Tend)

]

where n(i−1)T ,o (n(i−1)

T ,e ) is the T onset (end) position, according to

iteration (i − 1), n(i−1)T ,f (n(i−1)

T ,l ) is the location of the first (last)significant maximum modulus, associated with the T wave inw

(i−1)d,m [n], and sCSE(To n ) = sCSE(Te n d ) = 15.3 ms is the toler-

ance value given in [8] for T-wave end (no tolerance value forT-wave onset was given).

If n(i)T ,f (n(i)

T ,l) has equal or lower amplitude than

n(i−1)T ,f (n(i−1)

T ,l ), it is considered that the lead constructed at step(i) is not better for T onset (end) location than that constructedat step (i − 1), and applied the iteration STOP criterion of itemd). The iteration STOP criterion e) is applied if the locations intwo consecutive steps differ by less than two samples.

D. Validation

The evaluation of the automatic delineation strategies wasperformed over real files from the available manually annotatedECG databases (as the true onsets and end on clinical ECG sig-nals are unknown). Two standard databases have been repeatedlyused for evaluation of ECG delineation systems: the CommonStandards for Electrocardiography ML measurement database(CSEDB [9], 42 short signals in 15 leads at 500 Hz) and theQT database (QTDB [10], 105 files, 15 min long, in two leadsat 250 Hz). In CSEDB, manual annotations were made by fivecardiologists having in view all the available leads (ML-based);for a beat per file are provided median referee annotations af-ter an elaborated reviewing scheme of four rounds with outlierrejection, designed to reduce intra- and interobserver variabil-ity. QTDB included annotations from two cardiologists, butonly the first referee is considered here, who provided marksfor at least 30 beats per file (in a total of more than 3600annotated beats). More recently, the Physikalisch-TechnischeBundesanstalt (PTB) database (PTBDB, 549 files at 1000 Hz)has also been manually annotated in the context of the “Phy-sioNet/computers in cardiology challenge 2006: QT intervalmeasurement” competition [11]. A set of reference annotations

(one beat per file) was published in [12], consisting of SL man-ual annotations for QRS onset and T-wave end (based on leadII), done by four cardiologists and one biomedical engineer, anda median referee annotation (after outlier rejection). Accordingto [12], in more than 15% of the records in PTBDB, no T wavecould be definitely recognized in lead II, and, in such a case, thereferees were instructed to mark the T-wave end as a group atone of the leads where the T wave was better manifested.

The delineation system was validated over these threedatabases by comparing the automatic marks found with theprovided referee marks. It should be remarked that not all beatsare annotated, nor all waves and boundaries are given for eachannotated beat. In CSEDB, a total of 42 beats are partially an-notated, including 32 complex QRS onsets, and 26 ends and 27T-wave ends, with no T onsets. In QTDB, more than 3600 an-notated beats are provided; nevertheless, given marks also varyfrom beat to beat. In PTBDB, only the QRS onset and T-waveend are available, in one beat per file, and, in six of those, thefirst beat, which is not reliable for WT methods due to bordereffect, is annotated; thus, only 542 beats can be considered.

In CSEDB and PTBDB databases, three different VCG sys-tems were considered: lead set F—defined by recorded orthog-onal Frank leads (X,Y,Z); lead set M—defined by leads V5,aVF, and V2, a subset of three mutually orthogonal leads out ofthe standard 12-lead system; lead set D—defined by the syn-thesized orthogonal leads (X,Y,Z) from the 12-lead system, byusing the coefficients provided by the inverse Dower matrix [7].These VCG systems were chosen because they are defined bywell-known leads, which are likely to be familiar to clinicians,but any other combination of three orthogonal leads could beused instead, and other transformations from 12 to 3 leads ratherthan Dower matrix can be considered. The ML over the systemsF , M , and D was compared with SL delineation over each ofthe 15 recorded leads. Additionally, SL followed by the post-processing decision rules (SLR) described in Section II-B wasapplied over 12 leads, and also over the 3 leads in each lead set(F , M , and D), but with k = 0 (no protection rules at all).

Since QTDB includes only two leads, ML was applied usingloops in the 2-D plane instead of in the 3-D space. Regarding theorthogonality of the available leads, the QTDB was divided intofour subgroups: QTDB1 —7 records with orthogonal leads fromthe 12-lead standard system, in which the ML delineation canbe applied directly using a 2-D approach; QTDB2—57 recordswith no identified leads, here assumed to be orthogonal andtreated as the ones in QTDB1; QTDB3—34 records with noorthogonal and no parallel leads, which were orthogonalized byconstructing a new ECG lead orthogonal to one of the providedleads; QTDB4—7 records with parallel leads, which cannot beorthogonalized, and therefore, will not be considered for thevalidation (leads I and V5 from the 12-lead standard system).Additionally, a combined mark was obtained by choosing foreach fiducial point the location on the lead with less error (bestmark). Although this last approach cannot be considered as arule to apply in real practice where no reference marks exist,it is a reasonable way to compare the two SL annotation setswith the manual annotations that have been performed havingin view all available leads.



Fig. 5. Delineation results in CSEDB: comparison between ML over lead subsets F , M , and D (darker gray area), SL over each of the 15 available leads, andSLR over the standard 12 leads and over the three orthogonal leads in F , M , and D (lighter gray area). The symbol # denotes the number of TP out of 32 referencemarks provided (black) or after excluding extreme cases in each approach (gray); results fulfilling loose criterion are marked with × and fulfilling strict criterionare marked with a star.

Fig. 6. Delineation results in PTBDB: comparison between ML over lead subsets F , M and D (darker gray area), SL over each of the 15 available leads, andSLR over the standard 12 leads and over the three orthogonal leads in F , M , and D (lighter gray area). The symbol # denotes the number of TP out of 542reference marks provided (black) or after excluding extreme cases in each approach (gray); results fulfilling loose criterion are marked with × and those fulfillingstrict criterion are marked with a star.



The detection performance was evaluated by calculating thesensitivity S = 100 TP/(TP + FN), where TP is the numberof true positive detections and FN stands for the number offalse negative detections. For the CSEDB, S was not calculatedgiven the low number of annotated beats. The delineation error(ε) was taken as the automatically detected boundary minus therespective referee mark, and in each database, the mean (mε)and standard deviation (sε) of ε were evaluated across files,considering all TP detections. Additionally, for CSEDB andPTBDB, the aforementioned parameters were also calculatedafter the exclusion of the 10% most extreme cases, which arelikely to be outliers. Concerning QTDB, since several beats areannotated per file, the standard deviation of the error ε was firstcalculated across beats for each file and then averaged acrossrecord. Thus, for this database, sε stands for the mean standarddeviation across files.

According to the recommendations in [8], the standard de-viation of the differences from the reference should not exceedcertain limits. These values (2sCSE ) correspond to two standarddeviations of the differences between the median of the individ-ual readers manual annotations used for constructing CSEDBand were already referred to in this paper as they were used todefine search windows in the ML delineation system (Section II-C). In spite of being widely accepted as a way for defining atolerance for the automatic marks errors dispersion, it is notconsensual if an algorithm should accomplish the following.

1) Loose criterion: s < 2sCSE [5], [13]–[15].2) Strict criterion: s < sCSE [16], [17].

III. RESULTS

ML approach was applied to each of the three different VCGsystems considered (F , M , and D). SL delineation was per-formed over each of the 15 recorded leads and SL followed bythe postprocessing decision rules (SLR) was applied over 12leads, and also over the three leads in each lead set (F , M , andD). In Figs. 5 and 6, mε and sε in CSEDB and PTBDB, respec-tively, are plotted, both considering all true positive detectionsand after excluding the extreme cases. As expected, results fromSL denote a high dependency on the specific lead considered.From the strategies based on multiple leads, the SLR over 12leads is the one that achieves the best performance, especially forQRS boundaries, but with lower sensitivity for T end in PTBDB.SLR over three leads presents a high error dispersion both forQRS complex onset and T-wave end. ML delineation over threeorthogonal leads achieved in all datasets better results than mostisolated leads, though slightly worse than the best SL result. Inmost cases, it presented a reduced error dispersion compared toSLR over three leads. In particular, ML over lead set F outper-formed any SL-based delineation for T-end delineation.

For the sake of comparison, other published results also val-idated in CSEDB can be found in Table I. Note that the resultsreported in [5] and [16] used the described decision rules overthe 15 sets of marks, while no information was provided in [13]about which ML rules were used.

Two typical situations for QRS onset delineation are illus-trated in Fig. 7. In Fig. 7(a) is presented a case in which theML methodology fails, as two out of the three orthogonal leads

TABLE IOTHER PUBLISHED DELINEATION SYSTEMS WITH SELECTION RULES

VALIDATED OVER CSEDB

Fig. 7. Comparison between SL and ML delineation in the case of smallinitial waves. (a) Example in which the ML fails (i = 2). (b) Case in which MLperforms correctly (i = 3). Notation used is the same as in Fig. 4(a).

showed lower errors than the ML approach. This is due to an ini-tial positive wave in the QRS that is not detected in the last step ofthe iterative method, resulting from a less suited lead direction.This type of errors is, however, not systematic in the presenceof small Q/R waves. As a matter of fact, in Fig. 7(b), one can



TABLE IIQRS BOUNDARIES DELINEATION RESULTS IN QTDB. (a) QRS COMPLEX ONSET. (b) QRS COMPLEX END

TABLE IIIT-WAVE BOUNDARIES DELINEATION RESULTS IN QTDB. (a) T-WAVE ONSET. (b) T-WAVE END

find a similar case in which the ML achieves an improvementin the delineation, in spite of the noise contamination.

In Tables II and III are presented the values of S and mε ,sε in each QTDB subgroup, and the global results (all). It wasfound that a relative low number of extreme cases were caus-ing a large fraction of the global error. The exclusion of the10% more extreme files in each approach allowed a generalized

improvement in the errors dispersion, with bias increase in somecases.

IV. DISCUSSION

Globally, ML allowed an error dispersion similar to thatobtained using SLR over the 12 leads, including other pub-lished approaches (Table I). High sensitivity values were



found, resulting from the fact that only a very low num-ber of boundaries were considered to be inconsistent anddiscarded.

The automatic procedures are marking the QRS onset onCSEDB files later than the referees, as can be seen from thepositive bias found in most of the cases. This is likely to be dueto situations as the ones illustrated in Fig. 7(a). With respect tothe PTBDB files, it is known that the manual QRS onset (all) andT end (85%) reference marks are SL-based in lead II [12]. Thisis reflected in the lowest bias found for SL over lead II in T-waveend location. Either ML or SLR when compared with SL-basedreference marks (PTBDB) show a negative bias for QRS onsetand a positive bias for T-wave end. Clearly, this is in accordancewith the strategy of using information from multiple leads inorder to locate the earliest and the latest signs of ventricularactivity.

The ML over the VCG was able to provide, from only threeECG leads, boundary locations as stable as the ones providedby other methods using many more leads. It should be remarkedthat the proposed method requires the WT calculation of threeleads, with delineation procedures involving a variable numberof signals. Thus, even by considering fitting and projecting fea-tures, the ML strategy is usually more efficient than applyingSL to 12 leads, as the number of steps needed is not very high.The number of iterations was low, with mean values betweenthree and four iterations for QRS boundaries and less than twofor T boundaries. Thus, the results presented denote a clearperformance improvement.

Among the VCG systems considered, lead set F achievedthe best global performance. This is in accordance with theresults of the “PhysioNet/computers in cardiology challenge2006: QT interval measurement,” in which the QT interval ina representative beat per file was evaluated over the PTBDBfiles [4], [11], [18].

With respect to ML delineation using only two leads (QTDBfiles), global results are similar to the worse SL result for QRSboundaries, thus outperforming the best SL result for T-waveboundaries. Using a 2-D approach over two orthogonal leadsseems insufficient for QRS boundaries. The ML delineation ofthe T wave in files with two orthogonal (or orthogonalized)leads is more stable than using any of the two leads by them-selves. Further studies should be carried out to fully evaluatethe performance using only two leads.

The methods were validated over the available ML databaseswith reference annotations, among which only the CSEDB canbe considered to have truly ML-based annotations, with thereferee considering all leads during the annotation process. Inaddition to the fact that CSEDB includes just a very limitednumber of beats, the referees typically tend to base the markin a dominant clear lead, instead of using combination of them.To have a fairer evaluation of the proposed methods, truly MLmanual annotations should be obtained, for instance, by pre-senting the VCG loop to the referee. These kinds of annota-tions, which are not available today, will be better matched tothe physiological boundary and would allow a better quantifi-cation of the true improvement achieved by the proposed MLmethods.

V. CONCLUSION

A novel ML WT-based strategy for ECG boundaries delin-eation was proposed here and evaluated with respect to the QRSand T-wave boundaries. Using different leads is crucial to locatethe global waves boundaries, which can be imperceptible in aparticular lead. The automatic proposed ML approach allows todeal with multiple leads, taking advantage of their availabilityto further improve the delineation, by constructing a WT signalmore fitted for the specific boundary location. The ML systemprovided more robust and more accurate boundaries locationsthan any electrocardiographic lead by itself and outperformedstrategies based on rule selection after SL delineation.

Among VCG systems, the lead set F (directly recorded Frankleads) achieved the best global performance. The results alsoshowed that ML delineation with two orthogonal (or orthogo-nalized) leads, especially for the T wave, outperforms both SLresults, and thus, is better than any possible selection rule forT end delineation. Furthermore, the ML delineation strategiesdeveloped are general and can be applied to any orthogonallead set; the VCG systems considered here were chosen be-cause they are defined by well-known leads, which are likelyto be familiar to clinicians; however, any other combination ofthree orthogonal leads (real or derived) could be used instead.In particular, this method has also been partially validated overthe first three principal components, outperforming the resultsusing inverse Dower transformation [19]. Note that techniquesbased on principal or independent components analysis havebeen widely used, both for filtering and segmentation of ECGsignals [20]–[22].

As far as we know, this is the first delineation system that ex-plicitly constructs a new better suited ECG lead for that purpose,instead of locating the fiducial points in one available lead. Thissolves the problem of different latencies on the waves’ onsetsand ends found in different leads and combine the informationprovided by the multiple leads, taking advantage of their spacialdependency, and giving a unique annotation for the onset andend of the cardiac electric phenomena. Globally, results withinthe tolerance bounds were obtained for all boundaries, showingthat the proposed methods are quite robust against noise andmorphological variations.

REFERENCES

[1] C. Li, C. Zheng, and C. Tai, “Detection of ECG characteristic points usingwavelet transforms,” IEEE Trans. Biomed. Eng., vol. 42, no. 1, pp. 21–28,Jan. 1995.

[2] J. P. Martınez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna,“Wavelet-based ECG delineator: Evaluation on standard databases,” IEEETrans. Biomed. Eng., vol. 51, no. 4, pp. 570–581, Apr. 2004.

[3] J. Malmivuo and R. Plonsey. (1995). Bioelectromagnetism—Principlesand Applications of Bioelectric and Biomagnetic Fields. London,U.K.: Oxford Univ. Press [Online]. Available: http://butler.cc.tut.fi/malmivuo/bem/bembook/index.htm (accessed Dec. 2006).

[4] R. Almeida, J. P. Martınez, A. P. Rocha, P. Laguna, and S. Olmos, “Au-tomatic multilead VCG based approach for QT interval measurement,” inProc. Comput. Cardiol. Conf. 2006, vol. 33, pp. 369–372.

[5] P. Laguna, R. Jane, and P. Caminal, “Automatic detection of wave bound-aries in multilead ECG signals: Validation with the CSE database,” Com-put. Biomed. Res., vol. 27, no. 1, pp. 45–60, Feb. 1994.

[6] S. Van Huffel and J. Vandewalle, The Total Least Squares Problem: Com-putational Aspects and Analysis (Frontiers in Applied Mathematics), vol.9. Philadelphia, PA: SIAM, 1991.



[7] G. E. Dower, “The ECGD: A derivation of the ECG from VCG leads,” J.Electrocardiol., vol. 17, no. 2, pp. 189–191, 1984.

[8] The CSE Working Party, “Recommendations for measurement standardsin quantitative electrocardiography,” Eur. J. Heart, vol. 6, pp. 815–825,1985.

[9] J. L. Willems, P. Arnaud, J. H. van Bemmel, P. J. Bourdillon, R. Degani,B. Denis, I. Graham, F. M. Harms, P. W. Macfarlane, and G. Mazzoccaet al., “A reference data base for multilead electrocardiographic computermeasurement programs,” J. Amer. Coll. Cardiol., vol. 10, no. 6, pp. 1313–1321, 1987.

[10] P. Laguna, R. G. Mark, A. Goldberger, and G. B. Moody, “A databasefor evaluation of algorithms for measurement of QT and other waveformintervals in the ECG,” in Proc. Comput. Cardiol. 1997, pp. 673–676.

[11] G. B. Moody, H. Koch, and U. Steinhoff, “The physionet computers incardiology challenge 2006: QT interval measurement,” in Proc. Comput.Cardiol. Conf. 2006, vol. 33, pp. 313–316.

[12] I. Christov, I. Otsinsky, I. Simova, R. Prokopova, E. Trendafilova, andS. Naydenov, “Dataset of manually measured QT intervals in the electro-cardiogram,” Biomed. Eng. Online, vol. 31, no. 5, pp. 5–31, 2006.

[13] J. S. Sahambi, S. N. Tandon, and R. K. P. Bhatt, “Using wavelet transformfor ECG characterization,” IEEE Eng. Med. Biol. Mag., vol. 16, no. 1,pp. 77–83, Jan. 1997.

[14] H. J. L. M. Vullings, M. H. G. Verhaegen, and H. B. Verbruggen, “Auto-mated ECG segmentation with dynamic time warping,” in Proc. 20th Ann.Int. Conf. IEEE Eng. Med. Biol. Soc. Hong Kong: IEEE, 1998, pp. 163–166.

[15] J. A. Vila, Y. Gang, J. M. Presedo, M. Fernandez-Delgado, and M. Malik,“A new approach for TU complex characterization,” IEEE Trans. Biomed.Eng., vol. 47, no. 6, pp. 764–772, Jun. 2000.

[16] P. de Chazal and B. Celler, “Automatic measurement of the QRS onset andoffset in individual ECG leads,” in Proc. 18th Ann. Int. Conf. IEEE Eng.Med. Biol. Soc., Amsterdam, The Netherlands: IEEE, 1996, pp. 1399–1400.

[17] P. Strumillo, “Nested median filtering for detecting T-wave offset inECGs,” Electron. Lett., vol. 38, no. 14, pp. 682–683, Jul. 2002.

[18] J. P. Martınez, R. Almeida, A. P. Rocha, P. Laguna, and S. Olmos, “Stabilityof QT measurements in the PTB database depending on the selected lead,”in Proc. Comput. Cardiol. Conf. 2006, vol. 33, pp. 341–344.

[19] R. Almeida, J. P. Martınez, A. P. Rocha, and P. Laguna, “QRS com-plex boundaries location for multilead electrocardiogram,” in Proc. Com-put. Stat. 2008, Int. Assoc. Stat. Comput. Symp., Heidelberg, Germany:Physica-Verlag, 2008, pp. 447–454.

[20] R. Sameni, C. Jutten, and M. B. Shamsollahi, “What ICA provides forECG processing: Application to noninvasive fetal ECG extraction,” inProc. 2006 IEEE Int. Symp. Signal Process. Inf. Technol., pp. 656–661.

[21] T. He, G. Clifford, and L. Tarassenko, “Application of independent compo-nent analysis in removing artefacts from the electrocardiogram,” NeuralComput. Appl., vol. 15, no. 2, pp. 105–116, 2006.

[22] J. L. Willems, P. Arnaud, J. H. van Bemmel, P. J. Bourdillon, R. Degani,B. Denis, I. Graham, F. M. Harms, P. W. Macfarlane, G. Mazzocca,J. Meyer, and C. Zywietz, “Principal component analysis in ECG signalprocessing,” EURASIP J. Adv. Signal Process., vol. 2007, pp. 74580-1–74580-21, 2007.

Rute Almeida was born in Porto, Portugal, in 1979.She received mathematics applied to technology de-gree, in 2000, and the Ph.D. degree in applied math-ematics, in 2007, both from the Faculty of Sciences,University of Porto (FCUP), Porto, Portugal.

Between March and October 2001, she was withthe Autonomic Function Study Center, Hospital S.Joao, where she worked on methods for automaticdelineation of the ECG. From October 2001 toSeptember 2005, she was supported by an individ-ual grant from the Foundation for Science and Tech-

nology (Portugal) and the European Social Fund. Since 2005, she has been amember of the Centro de Matematica da Universidade do Porto (CMUP), Porto.Since 2007, she has been a Researcher at the Spanish Center for BiomedicalEngineering, Biomaterial and Nano-medicine Research Centro de InvestigacionBiomedica en Red en Bioingenierıa, Biomateriales y Nanomedicina (CIBER-BBN), under protocol with the University of Zaragoza (UNIZAR), Zaragoza,Spain. She is also with the Aragon Institute for Engineering Research (I3A)at UNIZAR. Her current research interests include time-scale methods and theautomatic analysis of the ECG, namely the study of ventricular repolarization.

Juan Pablo Martınez was born in Zaragoza, Spain,in 1976. He received the M.S. degree in telecommu-nication engineering and the Ph.D. degree in biomed-ical engineering from the University of Zaragoza(UZ), Zaragoza, in 1999 and 2005, respectively.

Since 2000, he has been an Assistant Profes-sor in the Department of Electronic Engineering andCommunications, UZ, where he has been an Asso-ciate Professor since 2007. He is a Researcher at theAragon Institute for Engineering Research (I3A), andis also with the Centro de Investigacion Biomedica

en Red en Bioingenierıa, Biomateriales y Nanomedicina (CIBER-BBN). Hiscurrent research interests include the field of biomedical signal processing, in-cluding signals of cardiovascular origin.

Ana Paula Rocha (M’05) was born in Coimbra,Portugal, in 1957. She received the Appl. Math. de-gree and the Ph.D. degree in applied mathematics,systems theory, and signal processing from the Fac-ulty of Sciences, University of Porto (FCUP), Porto,Portugal, in 1980 and 1993, respectively.

She is currently an Auxiliar Professor in the De-partment of Applied Mathematics, FCUP, where sheis also a member of the Centro de Matematica daUniversidade do Porto. Her current research inter-ests include biomedical signals and system analysis

(EMG, cardiovascular systems analysis, and autonomic nervous system charac-terization), time–frequency/time-scale signal analysis, point processes spectralanalysis, data treatment, and interpretation.

Pablo Laguna (M’92–SM’06) was born in Jaca,Spain, in 1962. He received the M.S. and Ph.D. de-grees in physics from the Science Faculty, Univer-sity of Zaragoza, Zaragoza, Spain, in 1985 and 1990,respectively.

From 1987 to 1992, he was an Assistant Profes-sor of automatic control in the Department of Con-trol Engineering, Politecnic University of Catalonia(UPC), Spain, where he was also a Researcher withthe Biomedical Engineering Division, Institute of Cy-bernetics (UPC-CSIC). From 1992 to 2005, he was

an Associated Professor at the University of Zaragoza, where he is a Full Pro-fessor of signal processing and communications in the Department of ElectricalEngineering, Engineering School, and also a Researcher at the Aragon Institutefor Engineering Research (I3A). He is also a member of the Spanish center forBiomedical Engineering, Biomaterial and Nanomedicine Research Centro deInvestigacion Biomedica en Red en Bioingenierıa, Biomateriales y Nanomedic-ina (CIBER-BBN). His current research interests include signal processing, inparticular applied to biomedical applications. He is a coauthor of BioelectricalSignal Processing in Cardiac and Neurological Applications (Elsevier, 2005).


Multilead ECG Delineation Using Spatially Projected Leads From Wavelet Transform Loops

Documents

Transcript of Multilead ECG Delineation Using Spatially Projected Leads From Wavelet Transform Loops